As the global population continues to aggregate in urban areas, numerous cities are rapidly growing along with substantial impacts on climate change, biodiversity, food production, and social-economic inequality. This makes it urgent to understand the changes in urban Land-use and Land-cover (LULC) patterns. At present, the understanding of the spatio-temporal evolution of LULC patterns is heavily constrained by the lack of efficient and precise methods of mapping and predicting LULC changes, particularly for the areas that have few established data sources. The main aim of the research is to fill this gap and develop a modelling framework to investigate urban LULC changes with the necessary efficiency and precision. The research makes use of publicly available multi-temporal Remote Sensing (RS) data. The use of such datasets to study large-scale urban regions is made feasible through the adoption of Deep Learning (DL)-based methods. The hypothesis of this study is DL methods could substantially facilitate the mapping and prediction of the changes in LULC patterns, particularly the transitions between LULC classes, by exploiting the spatio-temporal heterogeneity of time series RS data. To examine this hypothesis, this study proposes and tests a novel DL-based modelling framework, which consists of a LULC classification module and a LULC prediction module. The logic of the proposed framework is employing the classification module to generate multi-temporal LULC maps, which are then employed as the input for the prediction module to project changes in LULC patterns. In the classification module, a method of post-classification relearning with recurrent convolutional neural network models is developed for improving the accuracy of multi-temporal LULC classification. Also, a DL-based image super-resolution method is developed for contributing to accuracy gains of multi-temporal LULC classification. As for the prediction module, a tailored DL-based ensemble framework is proposed for LULC prediction, the proposed method adopts transformers as the base learners to handle spatio-temporal features and incorporates an attention mechanism to indicate feature importance. The proposed prediction method is also tested for simulating likely scenarios with different urban expansion rates. The research contributes to the advancement of knowledge about (i) exploiting state-of-the-art DL methods for mapping the spatio-temporal heterogeneity of LULC patterns at granular level, (ii) developing interpretable DL methods for indicating feature importance in LULC prediction, and (iii) integrating multiple DL methods into a modelling framework for LULC prediction based on RS data. Notably, although the research focuses on the patterns of LULC changes, the proposed framework can be generalizable and applicable in other studies associated with time series geospatial data.